Fig. 1: A schematic diagram of the key idea in this study.

(1) Our main task is to train a skill valuation model with machine learning technology. Under the paradigm of supervised learning, we need a set of training data with explicit labels of skill value to provide supervision for the model. Then the model can learn a function that maps the input (i.e., context and skills) to the observation (i.e., skill value). However, the labeled data of skill value is unavailable in our dataset. (2) We have abundant data of job postings with labels of salary, which can provide supervision for training a salary prediction model. Therefore, with the intuition that valuable skills should lead to high job salary, we regard salary prediction as a cooperative task that provides indirect supervision for skill valuation model. (3) We propose a model, SSCN, to simultaneously achieve skill valuation and salary prediction tasks, where the skill valuation model is a component of the salary prediction model. Specifically, SSCN estimates the skill value and composes skill value into job salary. In this way, the skill valuation model can be trained with feedbacks from the salary prediction task.